Replicable Bandits for Digital Health Interventions

Area of Science:

• Digital Health
• Clinical Trials
• Causal Inference

Background:

• Adaptive treatment assignment algorithms, like bandit algorithms, are prevalent in digital health clinical trials.
• Data from these trials often informs intervention refinement and broader deployment decisions.
• Inference for adaptive algorithm-dependent estimands, such as mean reward, is crucial but challenging.

Purpose of the Study:

• To investigate the replicability of statistical analyses in trials using adaptive treatment assignment.
• To identify why standard statistical estimators may fail in these adaptive settings.
• To introduce a formal definition and framework for replicable adaptive algorithms.

Main Methods:

• Theoretical analysis of statistical estimators under adaptive algorithms.
• Introduction and formal definition of "replicable bandit algorithms".
• Simulation studies using a mobile health oral health self-care intervention.

Main Results:

• Standard statistical estimators can be inconsistent and non-replicable in adaptive trials, even with large sample sizes.
• Non-replicability is intrinsically linked to the properties of the adaptive algorithm.
• Under "replicable bandit algorithms", common estimators are guaranteed to be consistent and asymptotically normal.

Conclusions:

• Designing adaptive algorithms with replicability is essential for reliable digital health interventions.
• Replicated evidence is critical for deployment decisions in digital health.
• Further research is needed on the interplay between adaptive algorithm design, statistical inference, and experimental replicability.